CN107851263B

CN107851263B - Method for processing recommendation request and recommendation engine

Info

Publication number: CN107851263B
Application number: CN201580081732.0A
Authority: CN
Inventors: 凯瑟琳·努赫拉·齐娜
Original assignee: B2yun
Current assignee: B2yun
Priority date: 2015-07-16
Filing date: 2015-07-16
Publication date: 2022-07-15
Anticipated expiration: 2035-07-16
Also published as: EP3323099A1; CN107851263A; WO2017009683A1

Abstract

The invention relates to a method for processing a recommendation request from a user of a device (1) connected to a recommendation platform (2), the recommendation platform (2) being further connected to at least one item server (3) referencing a plurality of items, to at least one internal rating server (4) and to at least one external rating server (5), the platform (2) comprising a storage unit (22) and a processing unit (21) storing a preference profile of said user, the method being characterized in that it comprises performing, by the processing unit (21), the steps of: (a) receiving a recommendation request comprising at least one attribute of a requested item, an identifier of a user and a length of time from a device (1); (b) selecting a subset of items referenced by the item server (3) that match the attributes and the user's preference profile; (c) determining an initial ordering of the items of the subset according to a preference profile of a user and scores provided by the internal rating server (4) and the external rating server (5); (d) dynamically updating the initial ordering; (e) when the time limit is over or when the user requests to interrupt the update, the current ranking is sent to the device (1) as the final ranking. The invention also relates to a system for carrying out said method.

Description

Method for processing recommendation request and recommendation engine

Technical Field

The field of the invention is that of recommendation engines.

More particularly, the present invention relates to a method for processing a recommendation request.

Background

The recommendation engine is a real-time information filtering system intended to automatically identify items/content that can interest a user, e.g. from the user's past selections.

Such systems are used particularly in a "business to consumer" (B2C) context, for example, to increase sales for e-commerce platforms.

The known recommendation engine therefore proposes:

-predicting a user's score for a given item;

-ordering the items for comparison thereof.

In such a B2C context, the most important feature of the recommendation engine is reactivity, i.e., the timeliness of the system in view of user preferences, interactions with the system, and so forth.

Such user-centric recommendation systems are referred to as "memory-based" because the ratings database is permanently maintained so as to be used directly to generate recommendations to active users. In contrast, in a "model-based" system, i.e., item-centric, data is pre-processed offline. When a service is run by a user, a "learning" model is used for prediction.

However, in a "business-to-business" (B2B) context, where the recommended recipients are not consumers, the requirements are quite different.

In particular, the engine should maximize the quality, confidence and transparency of recommendations over reactivity. Furthermore, the engine should ensure good resilience against noise, corrupted data and electronic spam.

B2B engines are known, for example, from european patent applications EP2466496 or EP 1014282.

However, known engines still require an improvement (in preference to reactivity) in the quality or reliability of their results, and suffer from a phenomenon known as "cold start": these engines require a huge amount of data to begin to perform. Furthermore, when only a few items have been rated by the user, the known recommendation engine is likely to use pseudo-ratings or default votes on the available items, which may result in distrust.

Accordingly, there is a strong need for an improved recommendation engine that overcomes these deficiencies.

Disclosure of Invention

For these purposes, the invention provides a method for processing a recommendation request from a user of a device connected to a recommendation platform over a public network, said recommendation platform further being connected over a secure network to at least one item server referencing a plurality of items and to at least one internal rating server, said recommendation platform further being connected over a public network to at least one external rating server, said platform comprising a storage unit and a processing unit storing a preference profile of said user,

said method is characterized in that it comprises the following steps performed by said processing unit:

(a) receiving the recommendation request including at least one attribute of a requested item, an identifier of the user, and a length of time from the device;

(b) selecting a subset of the items referenced by the item server that match the attributes and the preference profile of the user;

(c) determining an initial ordering of the items of the subset according to the preference profile of the user and scores provided by the internal rating server and the external rating server;

(d) dynamically updating the initial ranking so as to take into account the evolution of the preference profile of the user and the scores provided by the internal rating server (4) and the external rating server (5);

(e) sending the current ranking as the final ranking to the device when the time limit ends or when the user requests to interrupt the updating.

Preferred and non-limiting features of the invention are as follows:

● a plurality of internal and external rating servers are included, step (c) comprising calculating a plurality of rankings and aggregating the rankings;

● each of the ranks being associated with a weight representing the rank or providing a trust level of an internal or external rating server;

● each weight is provided by the internal or external rating server associated with the ranking of the weights;

● step (d) includes measuring, by the processing unit, the accuracy and/or recall of the ranking and dynamically updating the associated weights according to the measurements;

● calculating an ordering associated with an internal or external rating server includes determining a score for an item from a selected subset of rating data;

● determining that the score of an item from the selected subset of ratings data published by the external rating server (5) includes validating the data to discard data identified as spurious or corrupt;

● step (d) includes iteratively calculating a plurality of successive rankings, each successive ranking calculated from at least a previous ranking, the preference profile of the user and a score provided by the internal or external rating server;

● calculating the ranking further includes aggregating a plurality of rankings into the current ranking.

In a second aspect, the invention provides a recommendation platform connected to at least one device and at least one external rating server over a public network and further connected to at least one item server referencing a plurality of items over a secure network and to at least one internal rating server, the platform comprising a storage unit storing at least a preference profile of the user and a processing unit,

characterized in that the processing unit is configured to perform:

-a receiving module for receiving a recommendation request from the device, the request comprising at least one attribute of a requested item, an identifier of the user and a length of time;

-a selection module selecting a subset of the items referenced by the item server that match the attributes and the preference profile of the user;

-a determination module determining an initial ordering of the items of the subset according to the preference profile of the user and scores provided by the internal rating server and the external rating server;

-an update module that dynamically updates the initial ranking so as to take into account the evolution of the preference profile of the user and the scores provided by the internal rating server and the external rating server;

-a sending module sending a current ranking as a final ranking to the device when the time limit is over or when the user requests to interrupt the updating.

According to a third aspect, the invention proposes a system comprising a recommendation platform according to the second aspect and at least one device.

Drawings

The above and other objects, features and advantages of the present invention will become apparent from the following detailed description of illustrative embodiments thereof, which is to be read in connection with the accompanying drawings, wherein:

figure 1 represents a system for carrying out the method according to the invention;

figure 2 is a schematic diagram representing the steps of the method according to the invention.

Detailed Description

With reference to the figures, a method according to a possible embodiment of the invention will now be described.

Overview of the System

The current approach is aimed at providing the user with a "recommendation report" from a recommendation request, i.e. providing the user with a list of items that can meet the needs the user expresses through a search request. The current "item" can be anything that is a good, product, multimedia content, service, and generally a solution to a user's problem.

As depicted by fig. 1, the current method is performed by a platform 2 implementing a recommendation engine. The platform 2 comprises a processing unit 21, e.g. a processor on any other data processing device, and a storage unit 22 (memory, typically a hard disk drive) storing a database of, inter alia, user preference profiles.

The platform 2 is connected to a number of devices and servers over a secure network 20, preferably under internet protocol, in particular a virtual private network VPN, within a public IP network 10.

In particular, each user uses a device 1, typically a personal workstation, a smart phone or a tablet. Each device 1 comprises its own processing unit, human interface, HCI (typically screen, keyboard, mouse, etc.) and is connected to the platform 2 via a network 10.

At its first connection to platform 2, the user is invited to create an account and fills out a search preference profile using the HCI. The interface also enables the user to generate a recommendation request (see below) and be delivered with the results (the recommendation report). The preference profile may, for example, contain relevant preference scores, response time preferences, types of external ratings, internal ratings from an expert knowledge base, and the like.

Also connected to the platform 2 are at least one item server 3, at least one internal rating server 4, also referred to as knowledge server (both local servers connected to the platform 2 via a secure local area network 20), and at least one external rating server 5.

The item servers 3 are local servers, each referencing a number of items, i.e. candidates, that may be able to fit the user. They include a database of data qualities of items. Among them, there are "qualified terms", i.e., possible matching terms, as will be explained later. Item server 3 typically presents a product or service.

Each candidate item is defined by a plurality of properties that enable the items to be compared together. In particular, the recommendation report presents a (ranked) subset of the candidate items in response to a search request generated by a user requesting the items.

Each of the internal rating server 4 and the external rating server 5 is a data source from which the ranking of qualifying items can be determined. More specifically, these internal 4 and external 5 servers are a local server and a remote third party server, respectively, that provide various ratings/scores/rankings of some of the referenced items. The server 5 may be a server of a forum for consumers, a website for experts, a social network, etc. Note that the item server 3 can function as the internal rating server 4.

The processing unit 21 thus processes many heterogeneous sources to compute different orderings of qualifying items. Advantageously, these sources are generally classified into four groups:

s1 ═ qualifying content database (i.e. data from item server 3) and preference profile;

s2 ═ various rating, and performance key indicators (from the external rating server 5);

s3 — knowledge and rating of expert (from internal rating server 4);

s4-feedback of the user and interaction of the user (from the internal rating server 4 and the external rating server 5, such as a social network).

In particular, the ranking may be determined by the group of sources (i.e., the ranking by sets S1, S2, etc.) and/or by the sources (i.e., each internal or external rating server 4, 5).

Note that the scores may be provided directly by

servers

4 and 5 or inferred by platform 2 from data published by each of these servers. To this end, the platform 2 may use a tool capable of extracting rating data from the

servers

4 and 5. For example, the tools may perform text analysis to read the shortcomings or advantages of reviews and identified items, and thus they may set forth scores.

As will be explained, it is preferable to have many heterogeneous internal/external rating sources in order to reduce the impact of electronic spam, noise or corrupt data and to improve the robustness of the system.

Method

The steps of the current method are represented by fig. 2. In a first step (a), the user first generates and sends a recommendation request to the platform 2.

The recommendation request defines a requirement specifically expressed in natural language by the user. Alternatively, the search request may be represented as directly defining one or more characteristics (filtering) of the requested item. In any case, the search request should include at least one attribute of the item requested by the user. In addition, the search request includes a length of time, i.e., a duration of the time period.

In fact, as will be explained, the current recommendation method defines a "dynamic session" of predefined length (for example one week). The length of time is selected by the user along with the properties of the item he wants to find. As will be explained, this length of time defines the duration that the platform 2 has been free to process to provide the recommendation report. In other words, the processing unit 21 of the platform 2 processes the data during this length of time in order to refine the results and make them more accurate.

Because we are in the B2B context, the quality of the results is prioritized over reactivity, and this length of time is accepted by the user. Furthermore, the user can decide between reactivity and quality by varying the length of time: the longer the duration, the higher the quality.

Subsets of items

In a further step (b), the processing unit 21 selects a subset of the items referenced by the item server 3 that match the attributes and the user's preference profile, i.e. identifies eligible items among the candidate items. This step is particularly included in filtering where items that apparently do not fit the user are discarded.

Advantageously, the platform 2 applies predetermined eligibility rules to identify the qualifying items.

Initial ordering and sequential ordering

In a third step (c) processing unit 21 determines an initial ordering of the items of the subset according to the user's preference profile and the scores provided by the internal rating server 4 and the external rating server 5. By ranking, it is meant the following ranking of qualified items: from the highest score (i.e., the item is most recommended to the user) to the lowest score (i.e., the item is least recommended to the user).

This initial ranking is the first ranking, which is immediately available when the user wants, but is not the optimal ranking. The ranking may be obtained by known ranking algorithms.

In particular, ratings data from internal sources (preference profiles, knowledge bases) and external sources (external rating servers, user feedback) are preferably aggregated.

For example, a "supervised ensemble ordering" approach may be applied to merge the different obtained orderings. By using the collective ordering method, the processing unit 21 of the platform 2 processes different orderings defined on the same set of items and calculates an ordering with the least divergence from each of the input orderings.

Let R ═ { R1., Rn } be the set of n different orderings defined on the same set of items (i.e., Ri is the permutation defined on the set of items). The ensemble ordering method calculates a new permutation R from the set R, for example:

where d is a distance function defined in the ordering. The following Kendal-tau distance metric may be used:

d (R1, R2) | (x, y) s.t.r1(x) < R1(y) and R2(x) > R2(y) |

Each used source S1, S2, S3, S4 (or each internal or external rating server 4, 5) processed by the processing unit 21 of the platform 2 can thus be associated with a weight representing the ordered trust value provided by the source (and more generally the trust level of the source).

A weighted ensemble ordering method is then applied. Initially, all sources are assigned equal weights (or weights provided by the servers 4, 5). Periodically, the accuracy and recall of the ranking provided by each source is measured for a batch of recent requests. The accuracy (recall, respectively) is given by the ratio of true positives over the sum of true positives and false positives (false negatives, respectively). The batch size and the update period are automatically set by the processing unit 21.

The weight of each source is advantageously preferably updated according to the precision and recall it induces. The monitoring process allows for the detection of sources of ranking (excluding spurious external ratings) of electronic spam or ill-performing. The fact that a weighted ensemble ranking method is used allows the system to be less sensitive to errors induced by several ranking sources. This also enables the "cold start" phenomenon to be avoided, as the initial ordering can always be computed with enough feasible data from the knowledge base source.

In addition, the current approach gives the ability to always adjust the weight and ordering of the various data sources according to their integrity and to enhance engine scalability and reliability.

It is noted that the external rating is preferably also verified by the processing unit 21 in order to identify false users and false ratings. The ranking is calculated without regard to the suspect external rating.

As already explained, the ordering will be refined during the length of time of the input. To this end, in step (d), the initial ordering is dynamically updated, in other words a continuous ordering is calculated.

More specifically, the ranking evolves by considering the evolution of the item specifications (as provided by the item server 3), the evolution of the user preference profile, the evolution of the internal rating server 4 and the external rating server 5 (sources can be added or removed), and the iteration of the evolution of the data from these sources. In particular, as explained, the weights associated with the sources can vary over time. More specifically, step (d) advantageously comprises dynamically updating the weights and thus the plurality of rankings to aggregate.

The "latest" ranking is thus calculated from the previously calculated rankings, the user's preference profile and the scores provided by the internal rating server 4 and the external rating server 5. More specifically, this is not a simple re-calculation of the ordering, but an iterative calculation that depends on the ordering of the previous values.

For example, it appears that sources of bias (large amounts of spurious data, high variability in ratings) will have limited impact on additional ranking calculations. For this reason, their weight in the fusion is reduced.

Thus, the ordering slowly "converges" to an optimal value. Advantageously, the formulas and rules for calculating the ranking are first predetermined according to a model, and then inferred by automatic learning, so as not to limit it and gradually improve its quality.

The recalculation of the ordering may be performed periodically or at each evolution of the data quality library.

Recommendation report

When the time limit is over, the processing unit 21 interrupts the calculation of the ranking and prepares the recommendation report. More specifically, it sends the current ranking as the "final ranking" to the device 1 for output. The final ranking may be accompanied by a composite and/or graphical representation of the recommendation data. If the user is not satisfied with the results, the session may be extended for a new length of time.

It is further noted that at any time at the end of a given period, the user may request that the update be interrupted and the current ordering become the final ordering, even if such ordering can have been further refined.

In any case, the final report is sent to the user as a message such as email, SMS (short message service, etc.).

During the entire process (i.e., before the time period expires), "updated reports" may be sent to the user as desired by the user. Such a report is a notification containing the current ranking sent to the device 1. For example, it may be decided that a report be sent to the user to notify it each time the ranking changes (i.e., if at least one item has its ranking modified after an update). Thus, the user is aware of the gradual improvement of the recommendations he has requested.

Platform and system

According to a second aspect, the invention proposes a platform 2 for performing the method according to the first aspect.

As explained, the recommendation platform 2 is connected to at least one item server 3 referencing a plurality of items and to at least one internal rating server 4 over a secure network 20. The recommendation platform 2 is also connected to at least one device 1 and to at least one external rating server 5 via a public network 10.

The platform 2 comprises a storage unit 22 for storing at least preference profiles of users of the devices 1 and a processing unit 21.

The processing unit 21 is configured to perform:

-a receiving module for receiving a recommendation request from the device 1, the request comprising at least one attribute of the requested item, an identifier of the user and a length of time;

a selection module that selects a subset of items referenced by the item server 3 that match the attributes and the user's preference profile;

-a determination module determining an initial ordering of the items of said subset according to the preference profile of the user and the scores provided by the internal rating server 4 and/or external rating server 5;

-an update module that dynamically updates the initial ordering;

-a sending module which sends the current ranking as the final ranking to the device 1 when said time limit is over or when the user requests to interrupt said updating.

In the case of multiple sort aggregates, the processing device may also execute a supervisor module that monitors the weights of the sources/sorts as explained.

Claims

1. A method for handling a recommendation request from a user of a device (1) connected to a recommendation platform (2) over a public network (10), the recommendation platform (2) further being connected to at least one item server (3) referencing a plurality of items and to at least one internal rating server (4) over a secure network (20), the recommendation platform (2) further being connected to at least one external rating server (5) over the public network (10), the platform (2) comprising a storage unit (22) and a processing unit (21) storing a preference profile of the user,

the method is characterized in that it comprises the execution, by said processing unit (21), of the following steps:

(a) receiving from the device (1) the recommendation request comprising at least one attribute of a requested item, an identifier of the user and a length of time defining a duration of time the recommendation platform has been free to process to provide a recommendation report;

(b) selecting a subset of the items referenced by the item server (3) that match the attributes and the preference profile of the user;

(c) determining an initial ranking of the items of the subset according to the preference profile of the user and scores provided by the internal rating server (4) and the external rating server (5), each of the internal rating server (4) and the external rating server (5) being a data source according to which a ranking is determined, each data source being provided by a respective rating/ranking;

(d) dynamically updating the initial ranking during the length of time, the updating comprising iteratively calculating a plurality of successive rankings, each successive ranking being calculated according to at least a previous ranking, the preference profile of the user and the evolution of the score provided by the internal rating server (4) and the external rating server (5), wherein step (d) comprises measuring, by the processing unit (21), the accuracy and/or recall of a ranking, and dynamically updating the associated weights according to the measurements;

(e) -sending the current ranking as the final ranking to the device (1) when the length of time is over or when the user requests to interrupt the updating.

2. A method according to claim 1, wherein a plurality of internal and external rating servers (4, 5) are included, step (c) comprising calculating a plurality of rankings and aggregating the rankings.

3. The method of claim 2, wherein each of the aggregated rankings is associated with a weight representing a trust level of the ranking.

4. A method according to claim 3, wherein each weight is provided by the internal or external rating server (4, 5) associated with the ranking of the weights.

5. A method according to claim 2, wherein calculating the ranking associated with an internal or external rating server (4, 5) comprises determining a score for an item from a selected subset of rating data.

6. A method according to claim 5, wherein determining scores for items from a selected subset of ratings data published by an external rating server (5) comprises validating the data so as to discard data identified as spurious or corrupt.

7. The method of any of claims 1-6, wherein calculating a rank further comprises aggregating a plurality of ranks into the current rank.

8. A recommendation platform (2), the recommendation platform (2) being connected to at least one device (1) and at least one external rating server (5) over a public network (10) and further being connected to at least one item server (3) referring to a plurality of items over a secure network (20) and to at least one internal rating server (4), the platform (2) comprising a storage unit (22) and a processing unit (21) storing at least a preference profile of the user,

characterized in that the processing unit (21) is configured to perform the method of any one of claims 1 to 6.

9. A recommendation system comprising a recommendation platform (2) according to claim 8 and at least one device (1).